Tibetan Ancient Manuscript Layout Analysis Based on Mcross-wmBiFPN and Hybrid Attention Mechanism

Tibetan ancient texts are invaluable treasures of Chinese cultural heritage, offering significant insights into Tibetan history, language, and culture. In this sense, the image-based layout analysis for Tibetan manuscripts is crucial in that it directly impacts the effectiveness of subsequent Tibetan text recognition. This paper focuses on the complex layout of Dunhuang Tibetan manuscripts and proposes an image layout analysis method based on Mcross-wmBiFPN and a hybrid attention mechanism. This method incorporates a hybrid attention mechanism to evaluate the importance of features, which improves the representational capability of the multi-scale feature pyramid. Additionally, by integrating cross-level connections across different layers, the method mitigates the potential information loss that may occur after multiple cascades of BiFPN. On the constructed dataset of Dunhuang Tibetan manuscript images, the proposed method achieved a mean Average Precision (mAP) of 99.17%, i.e. a 6.14% improvement over the baseline model EfficientDet.